Carlos E. Perez (@IntuitMachine) header
Carlos E. Perez (@IntuitMachine) profile picture

Quaternion Process Theory, Artificial (Intuition, Fluency, Empathy), Patterns for (Gen, LRM, Agentic, Skill) AI, https://teapawt.net/t.co/Ei2M8F8QOs

Arlington, VA
Joined February 2015
Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)Photo by Carlos E. Perez (@IntuitMachine)
My Artificial Cognition trilogy is now available for free. I began writing the first book a decade ago (2016) and completed the last two in 2023. The process of writing was an aid to understanding the development of today's AI. Insights in these books have stood the test of time and offer powerful metaphors for understanding today's AI and its future evolution. Feel free to have your daily AI driver ingest them for additional insight!
5
9
1
60
15,734
OpenAI lacluster 6.2 release bumped up the price!!
BREAKING: Anthropic, $ANT, is planning to raise its IPO price after stronger than expected investor demand, per WSJ. The company could now price its shares over $80 per share, above the initial range of $70 to $78 per share. Its advisers have discussed putting its valuation at around $1.5 trillion or $1.6 trillion, up from the ~$1.33 trillion valuation of the existing range. Demand appears to be strong for Anthropics IPO.
7
1,494
Yet another upgrade in the middle of the day.
The NSA system outage was actually just a software update -WSJ
1
15
1,539
Likely internal patching
Having spent time at NSA with deep knowledge of these systems, this one is different. NSANET is a true standalone, air-gapped network… physically isolated by design, with no external connections. That’s the entire point of the top-secret infrastructure. An outage there, with no indication of outside access, points to something internal. Speculating: hardware failure, a bad config change, power/cooling issue, or an insider problem. Official “no cyberattack” language doesn’t rule those out, and it doesn’t explain why a fully isolated network just went dark.
1
412
Kind of makes sense of why it's in the middle of the day!
A U.S. official familiar with the matter said “routine patching of systems” caused the NSA’s internal communication network to go down. “There’s just a lot more patching happening these days because AI is identifying more vulnerabilities,” the official said. “And the volume can cause networks to go down.”
2
2
7
1,269
Mythos likely working overtime!
Anne Neuberger worked at the NSA for over a decade focusing on cyber security. Her update on what’s been going on, particularly at the NSA, is well worth your time.
1
350
@grok when did Anthropic announce Mythos 5.2 and Fable 5.2?
1
1
237
1/3 Most AI tools give you another tab to babysit. Spara just shipped Scout—the first AI sales assistant that lives in iMessage & Slack and works your pipeline before you wake up. 2/3 — The Core Shift • Proactive, not prompted: Scans your CRM, calls, inbox & LinkedIn 24/7 for 6 AM job changes and dropped follow-ups. • Zero tab fatigue: Drops ready-to-approve drafts straight into Gmail & LinkedIn DMs. • Adapts to you: Learns how you sell—no rigid playbooks. 3/3 — The Proof Built by repeat founders David Walker (350-person GTM exit) & Zander Pease (Nomad Health). Early numbers: 3x reply rates (10.5% vs. 3.5%) and 29% more meetings booked. Watch the launch video above and test Scout. 👇
🤝 Paid partnership
Today we’re launching Scout, the AI assistant built for sales. It lives in your Slack and text. It works every account in your book and tells you who to talk to, why now, and what to say. Then learns how you sell and drafts the next best action to take in your voice: an email, a LinkedIn message ready to send. We ran it on our own outbound first. Reply rates went up 10X. The future of sales is here, and it’s about to close you a lot of deals. Request early access: spara.com/scout?utm_source=x
2
702
At least they aren't using Claude internally. ;-)
The highly classified network that the National Security Agency uses for its internal communications is down, according to six people familiar with the matter. At this point there is no evidence of a cyberattack, sources said. CNN has asked the National Security Agency for comment. cnn.it/4AA0eTd
1
1
7
1,144
Very concerning that this is an internal network.
220
Why would firms perform upgrades in the middle of the workday?
Markwayne Mullin: "It appears the technology that's happening with Microsoft or with Chase, we saw it with United Airlines, some of the technology that they are pushing out, they are pushing new technology on top of really old computer systems and it just seems to overwhelm the system a little bit."
1
1
3
1,113
1/ 90% of the Claude Code team's code is now written *and* reviewed by Claude. Without strong verification, you can only run 1-2 agents at a time. With it, you can run dozens or hundreds in parallel. This is the unlock. 🧵 2/ Verification loops are what let you actually delegate. In internal tests, having a separate grader check Claude's work against a rubric improved success rates by up to 10 points — biggest gains on the hardest problems. Evidence > vibes. 3/ The old model of verification was expensive: humans writing tests, running benchmarks, checking specs. Now agents can generate those tests/benchmarks themselves, and once written they run as normal software (cheap, frequent, no extra tokens). The cost curve flipped. 4/ Not all tests are equal. The Claude Code team prioritizes three types: • End-to-end (E2E) • Property • Differential These are the ones worth investing in. Everything else is secondary. 5/ E2E tests are the highest ROI. They describe real user flows. Either the feature works or it doesn't. Hard to game, clear done criteria, and they surface what's actually broken (screenshots + context beat "timeout after 30s"). Unit tests can all pass while the product is broken. 6/ Property tests catch what humans (and agents) miss. You write an invariant you believe can't fail ("balance never goes negative", "encode → decode = original", "operation is idempotent"), then let the library throw thousands of inputs at it. Agents are biased. Property tests aren't. 7/ Differential tests are the migration superpower. Run old system and new system on the same inputs (random or production traffic). Differences = bugs or intentional breaks. Especially powerful when the spec has drifted from reality. 8/ How the Claude Code team actually does it: They give Claude "hands and eyes" on the real running system via a `verify` skill (Chrome extension, mobile simulators, local services, computer use). Verification means *running the change*, not just checking boxes. 9/ Other hard rules they follow: • Failing checks get 3 auto-fix attempts, then a human • Code owners write their own checks • Changes cannot modify the checks that judge them (read from base branch) • Start new gates in "observe" mode before enforcing • Turn misses into explicit rules, then prune them regularly 10/ The end state: Your job shifts from reading every diff to defining what "done" looks like and reviewing the evidence. That's what lets you scale from a couple of agents to a fleet. Verification isn't optional anymore — it's the multiplier. Source: claude.dev/blog/how-the-clau…
10
1
12
1,351
Property-based tests generated from the spec and from the code catch different bugs. Use both. In a new paper, across 25 agent-built apps, Spec-Driven PBT found 56 bugs the Code-Driven tests missed, and Code-Driven PBT found 43 that Spec-Driven missed, so 59% of bugs were found by only one approach. Setup - An agent writes invariants in plain English from the requirements doc - Then it reads the code to turn them into property tests - The baseline derives properties from the code alone Why PBT? - Each property-based test kills about 50× as many mutants as the average unit test (prior work cited in paper) - Properties double as documentation, far shorter than the implementation Bugs only one approach found, by type - Spec-Driven: mostly state, lifecycle and logic errors (61% of 56) - Code-Driven: mostly input handling (49% of 43) Limitations - Bugs were judged by the same model that wrote the tests (Claude Opus 5); 25 sampled by hand, all held up - The prompts differed between approaches - The method needs a good spec Insight "Invariants obtained from the specification are not biased by the behaviour of a particular implementation, and can therefore falsify behaviour about which the implementation is internally consistent." My take The code can’t be the only oracle, the spec must be one too. Next: move key properties into types and checks, so they hold by construction. arxiv.org/abs/2611.15920
2
247
A new AI harness that has been fine-tuned for banking. Reluability in reasoning and accuracy are non-negotiable. We will see more specialized harnesses in the future.
Every bank has people capable of extraordinary work. Yet most of them are stuck navigating painful old systems. Those same systems are what hold AI back. On a modern stack, AI can do real work, and @StitchHQ now has it built in: - A copilot that does the work, not just the talking. - A banking-grade harness to natively embed your own agents. - Ready-made agents that drop into your existing workflows. All within the limits your bank sets. So your team gets to do the work it's capable of, with the tools they deserve. PS: I heard central banks now ask Jeff to audit them.
2
2
844
🚨 Google just confirmed the first fully autonomous AI cyberattack. No human in the loop. At all. A Russian-speaking crew’s “exploit_agent” read patches, wrote exploits, and ran them itself. 22 zero-days in August alone. 🧵 --- 1/ It’s a multi-agent system that: → Diffs code before & after a patch → Spots the bug → Writes the exploit → Fires it automatically First time Google has seen the full chain run without a human. --- 2/ Found on an exposed server next to stolen data + scripts for: - Recon - Password brute-forcing - Dropping webshells - Even deleting rival attackers’ webshells Russian comments in the config. Targeted French media, entertainment & education sites. --- 3/ The speed is the real problem. These tools trade stealth for velocity. Patch windows are collapsing. Elite exploit writers are no longer required — just API keys. --- 4/ Zero-days spiked to 22 in August (from ~8/month in 2025). Agentic attacks just crossed the threshold security teams have been dreading. Every company online just got a shorter fuse. --- 5/ Source: Google Threat Intelligence Group AI Threat Tracker (Nov 16, 2026) “Agentic Threat Activity Crosses a Significant Threshold” Thoughts?
10
16
7
87
7,466
Nice to know when to choose Argon vs Sol!
In AA-PersonalityBench, Gemini 4 Argon is the most metaphorical of all models measured, followed by Gemini 3.8 Flash and Muse Spark Open The metaphorical trait measures whether a model uses analogies or figurative language to make ideas tangible. It is one of five traits in the Expressiveness index. ➤ Most metaphorical: Gemini 4 Argon (100), Gemini 3.8 Flash (98), Muse Spark Open (95), Gemini 3.1 Pro Preview (93) and Inkling (91) ➤ Least metaphorical: GPT-3.5 Turbo (1), GPT-6 Luna (3), GPT-6 Sol (6), Llama 4 Maverick (8) and GPT-4o (Nov) (10)
3
854
1/ Anthropic’s Founder playbook is blunt: the traditional startup arc is dead. Validate → raise → hire → build → raise again is obsolete. AI-native founders ship production software and reach revenue with tiny teams. 2/ Non-technical domain experts can now build what used to require a technical co-founder. The wall between “people who can build” and “people with ideas” is gone. Founders are orchestrators of agents, not individual contributors. 3/ Idea stage: Stop coding until the evidence is real. Use AI for research, interview design, and synthesis. Hunch-based building is amateur hour. Exit only with problem-solution fit. 4/ MVP stage: Ship fast with agentic coding. Set retention and activation metrics *before* launch. The Sean Ellis test still matters. Pivot when the data says so—AI makes the diagnostic cheap. 5/ Launch & Scale: Turn traction into a repeatable engine. Codify knowledge, automate ops, build moats through depth and integrations—not headcount. Organic founder-led growth has a ceiling; AI lets you bootstrap GTM and systems without the old org chart. 6/ Same job, new rules. Taste, judgment, and facing disconfirming evidence still matter. Everything else—code, research, ops, analysis—can be orchestrated. Leverage > headcount. The shortest path from idea to exit just got dramatically shorter.
12
2
22
1,429